I’m Camille. Last week I was scrolling through the Artificial Analysis Video Arena leaderboard — the one where real users vote on blind video comparisons — and a name I’d never seen was sitting at #1. HappyHorse-1.0. No team page. No brand. GitHub links that say “coming soon.” If you make product videos or visual content with AI tools, a pseudonymous model topping both text-to-video and image-to-video rankings is worth understanding before the hype cycle decides for you.
This is what I’ve been able to confirm, what stays unverified, and why — if you care about input-driven video workflows — the I2V numbers deserve a closer look.
What Artificial Analysis Has Confirmed About HappyHorse-1.0
Let me start with the one hard signal we actually have. On April 7, 2026, Artificial Analysis posted on X that they had added a new model to their Video Arena. Their exact word: “pseudonymous.” That’s the entire confirmed identity.
“Pseudonymous model” — what that label means
Pseudonymous means the model was submitted without a verifiable team attached. It showed up, generated outputs, users voted blind — same as every other model in the arena. Anonymous benchmark drops have happened before, but one landing at #1 across multiple categories is unusual.
Current rankings across four leaderboards (T2V/I2V, with/without audio)
As of early April 2026, HappyHorse-1.0 holds these positions on the Artificial Analysis leaderboards:
| Category | Rank | Elo Score | Runner-Up (Elo) |
| T2V, no audio | #1 | 1,333 | Seedance 2.0 (1,273) |
| I2V, no audio | #1 | 1,392 | Seedance 2.0 (1,355) |
| T2V, with audio | #2 | 1,205 | Seedance 2.0 #1 (1,219) |
| I2V, with audio | #2 | 1,161 | Seedance 2.0 #1 (1,162) |
A 60-point Elo gap (T2V no audio) means one model wins roughly 58–59% of blind matchups — meaningful. A 1-point gap (I2V with audio) is noise. Newly added models tend to be more volatile than established ones with thousands of votes behind them. Check the live leaderboard before making decisions — these numbers will have moved by the time you read this.
What Multiple HappyHorse Sites Claim — and What Can’t Be Verified Yet
Here’s where things get murky.
15B parameters, open source, 38s inference, 7-language lip-sync
Several sites — happyhorse-ai.com, happy-horse.art, happyhorse.app, happy-horse.net, happyhorseai.net — each describe the same model: 15-billion-parameter single-stream Transformer, 40 layers, joint video-and-audio generation, 7-language lip-sync (Mandarin, Cantonese, English, Japanese, Korean, German, French), ~38 seconds for 1080p on a single H100, full commercial license.
These are claimed specs. I can’t independently verify any of them.
GitHub and Hugging Face links: status as of publish date
As of April 8, 2026, the GitHub and Hugging Face links on these HappyHorse sites point to “coming soon” pages or return 404 errors. The weights aren’t publicly downloadable.
Here’s what makes this interesting: a separate open-source project called daVinci-MagiHuman, developed by Sand.ai and GAIR Lab, is publicly available under Apache 2.0 — and it matches HappyHorse’s claimed specs almost exactly. Same parameter count, same architecture, same language list, same inference speeds. A 36Kr investigation found the benchmark numbers and website structures to be near-identical.
Team identity — community speculation vs confirmed facts
Nobody has officially claimed HappyHorse-1.0. Speculation on X has pointed at WAN 2.7, DeepSeek, Tencent, and — most persistently — Sand.ai’s daVinci-MagiHuman. The Year of the Horse timing (2026 in the Chinese lunar calendar) and language ordering on the sites (Mandarin before English) suggest an Asia-based origin.
The prevailing theory, per 36Kr, is that HappyHorse is an optimized daVinci-MagiHuman iteration submitted to stress-test user preference. But theory is not confirmation.
Why Image-to-Video Users Should Pay Attention
This is the part I care about for my own work.
Elo 1392 in I2V (no audio) — what strong reference-following means for input-driven workflows
The I2V no-audio score of 1,392 is the highest on the board. What that tells us: when users upload a reference image and compare blind results, HappyHorse’s output wins more often. The model appears to follow the reference more closely — subject identity, composition, visual coherence.
If you’re doing product videos or brand content where you start with a specific image and need the motion to respect that image, reference-following is the metric that matters most. Beautiful motion that drifts from your product shape isn’t useful. Locked-on motion that moves your subject convincingly is.
How clean cutouts and asset quality change AI video output
This applies regardless of which model you’re running: input image quality determines the ceiling of your output video. True for Seedance 2.0, true for Kling, and it’ll be true for HappyHorse if and when it becomes accessible.
Dirty edges, leftover halos, compression artifacts — the model reads all of that as signal and amplifies it into motion. I’ve covered this in detail for Seedance 2.0 asset prep and flicker edge cleanup. The logic is identical for any I2V tool: ghost halo around the cap in your photo means shimmering ghost halo in every frame of your video.
Before you rewrite a prompt, check the asset.
Known Limits and Open Questions
You can’t use it yet. No public API, no downloadable weights under the HappyHorse name. If daVinci-MagiHuman is the base model, that is available — but it requires H100-class hardware and nontrivial setup.
Elo scores are early. Seedance 2.0 has 7,500+ vote samples in T2V; HappyHorse’s count isn’t broken out. More votes could shift rankings either way.
Community testing is mixed. Some users on X report gaps with Seedance 2.0 in character detail and dynamic coherence. Others are excited about multi-shot potential. Short blind clips may not reflect all use cases.
Portrait-heavy evaluation. Per 36Kr’s analysis, portrait and voice-over content accounts for 60%+ of the arena’s test samples — giving face-and-speech-optimized models a built-in advantage.
FAQ
Is HappyHorse-1.0 made by ByteDance / DeepSeek / Tencent?
No official confirmation for any of these. The most discussed theory links it to Sand.ai’s daVinci-MagiHuman, but that remains unconfirmed. Community guesses are not identification.
Can I use HappyHorse-1.0 for commercial video?
Not directly — there’s no public access under the HappyHorse name as of April 2026. Third-party demo sites offer browser-based generation with their own terms, but they’re not the model developer. If the underlying model is daVinci-MagiHuman, its weights are under Apache 2.0 (commercial use permitted) — but you’d need H100-class hardware.
Does input image quality affect HappyHorse output?
Yes — every I2V model amplifies input flaws into motion flicker. Cleaning your cutouts first saves more time than rewriting prompts.
How often do Artificial Analysis Elo scores change?
Continuously. A model at #1 today might be #3 next week. Always check the live leaderboard rather than relying on any article — including this one.
Alright, that’s where things stand. A model nobody can name is at the top of the most credible video benchmark we have, and the gap between confirmed and claimed is wide. My take: pay attention, but don’t rearrange your workflow around a model you can’t access yet. The thing you can control is how clean your input assets are — and that matters no matter which model wins.
See you next time.
Previous posts:
Seedance 2.0 Audio Guide: Dialogue, SFX, BGM, and Lip Sync Tips
AI Image to Video Online: Turn Any Photo Into a Motion Clip (Free)
Seedance 2.0 Image to Video: Turn One Photo Into a Consistent 16s Clip